Chinese AI models take all four top usage spots

Developers pushed 69 trillion tokens through AI models worldwide between August 3 and August 9. DeepSeek-V4-Flash-0731, released officially on July 31, took the top spot with 8.83 trillion tokens, after usage rose 570 percent week-on-week. The result was visible almost immediately in the global traffic chart.

Chinese models accounted for 34.25 trillion tokens, compared with 9.17 trillion for US models, according to the latest OpenRouter dataset. Chinese models have now led the global count for fifteen consecutive weeks. Tencent Hy3 ranked second with 8.05 trillion tokens, followed by DeepSeek’s preview version at 5.88 trillion and Xiaomi MiMo-V2.5 at 5.39 trillion. All four leading positions went to Chinese labs.

The model itself was not rebuilt from the ground up. Its structure and parameter count stayed unchanged, while DeepSeek reran the post-training recipe from scratch. The source links the resulting improvement in math and code benchmarks to the rapid shift in developer traffic. Artificial Analysis put the official V4-Flash’s average cost for a single-task benchmark at 0.03 dollars, roughly one one-hundredth of Anthropic’s Claude Fable 5 flagship, and described it as the lowest-cost major model in the world.

The ranking remains fluid. OpenAI’s GPT-5.6 Luna entered the top five at 4.43 trillion tokens after an 80 percent enterprise price cut, while Google’s Gemini 3.6 Flash entered the top ten for the first time at 2.33 trillion tokens, up 446 percent week-on-week. MiniMax M3 and Stepfun Step 3.7 Flash fell out of the top ten.

So what does it change in practice? For developers choosing a model at scale, the data points to Chinese open-source systems competing not only through lower prices but also through usage, throughput and benchmark performance. The numbers show a strong shift in usage, but they are still a weekly OpenRouter snapshot; the cost comparison comes from Artificial Analysis.